Why “AI-proofing” assessments misses the point—and what to design for instead

There is a category of business decision that feels like action but isn’t. You know it when you see it. The strategy offsite that produces a deck but no accountability. The risk framework that gets updated after every incident but never prevents the next one. The compliance programme that keeps the auditors satisfied while the underlying behaviour continues unchanged.

“AI-proofing” assessment is that decision.

I don’t say that to dismiss the instinct behind it. When a new technology appears to threaten the integrity of something you care about, the natural response is to design the threat out. Lock it down. Set the rules. Make the problem go away. In most operational contexts, that instinct is right.

In this one, it isn’t. And the reason why tells you something important about what actually needs to happen.

The wrong assumption underneath the wrong goal

AI-proofing rests on two assumptions that don’t hold.

The first is that AI is the problem. It isn’t. AI is a capability that has changed the conditions under which students complete assessments. The problem is assessment design that no longer produces reliable evidence of learning under those conditions. That’s a design problem. Targeting AI is targeting the symptom.

The second assumption is that removal is possible. It isn’t. AI is ambient. It is already embedded in how students write, research, think, and work—and it will become more so. Every field these graduates are entering is already using it. I talk to employers regularly. Not one of them is asking for graduates who don’t know how to use AI. They are asking for graduates who can think clearly, exercise judgment, and produce work they can stand behind. Those are different requirements— and they have significant implications for what assessment should be testing.

Designing assessments that assume AI doesn’t exist is not a long-term position. It is a temporary friction that decays the moment students work out how to route around it, which they will, because they’re rational and the system is giving them no reason not to.

What AI-proofing actually produces

Here’s what I see when institutions pursue this path, and I want to name it directly because I think it’s underacknowledged.

AI-proofing produces poor design behaviours. Over-constraining tasks. Surveillance-heavy formats — proctoring, lockdown browsers, in-person conditions applied to tasks that didn’t previously require them. Designing for detection rather than for demonstration.

The effort goes up. The educational quality doesn’t.

But here’s the deeper problem that doesn’t show up in the integrity data: when your assessment system is primarily organised around restriction, you are producing graduates who have been trained to navigate constraints rather than develop capability. The student who has spent three years learning to work around AI restrictions has not learned how to work effectively with AI. The student who has spent three years writing generic essays under controlled conditions has not learned how to communicate complex ideas in the environments they’ll actually encounter.

Those are the capabilities that high-stakes, output-only assessment has always struggled to develop. AI has just made that failure more visible.

Students are not confused. They’re rational.

Students are not behaving badly. They are behaving rationally inside the system they can see. If the system tells them—through its design, its incentives, its signals—that what matters is the output, they will optimise for the output. If the system tells them they will need to demonstrate their understanding later, they will prepare to demonstrate their understanding. The behaviour follows the design.

Research from Wonkhe makes this concrete: a single structural factor—whether students expect to be asked to explain their work—can fundamentally change how they use AI. Not whether they use it. How. When demonstration is required, AI becomes a tool in the thinking process. When it isn’t, AI can become a substitute for it.

That finding should be on the wall of every curriculum committee in the country. The leverage point is not the AI. It is the expectation of accountability. And that expectation is entirely within the institution’s control to set.

What to design for instead

The practical shift is less dramatic than it might seem. This doesn’t require pulling everything up from the roots and starting again. It requires intentional changes to task design that shift what the system rewards—and in doing so, align assessment far more directly with the capabilities graduates actually need.

Five levers, applied deliberately, change behaviour significantly.

Accountability moments.

Not necessarily formal examinations. But designed moments where students are expected to explain, justify, or respond to questions about their own work. A ten-minute conversation. A brief oral defence. A checkpoint where the student walks through their reasoning. These moments cannot be outsourced—and they directly develop the skill that comes up in almost every conversation I have with employers: the ability to communicate your thinking clearly and defend it under pressure. That is not a nice-to-have in any professional environment. It is table stakes.

Process capture.

Making the journey visible, not just the destination. Drafts, checkpoints, documented iterations. Beyond its value as an integrity mechanism, this builds something that the modern workplace demands constantly: the ability to manage a piece of work through multiple stages, incorporate feedback, and make visible decisions at each step. Consulting, law, engineering, medicine, financial services—every knowledge profession runs on iteration. Assessment that only evaluates the final product is producing graduates who have never been asked to show their working.

Low-stakes explanation.

Conversation over performance. The formal high-stakes submission creates maximum pressure and minimum insight into the thinking behind it. Embedding lower-stakes moments of explanation—in tutorials, in feedback sessions, in brief check-ins—builds a picture of genuine understanding. It also builds something that cannot be taught in a classroom but can be developed through practice: the ability to think on your feet, articulate ideas without a script, and engage with challenge in real time. These are the skills that separate good graduates from exceptional ones in every hiring conversation I’ve been part of.

Task specificity.

Reducing the generic, AI-friendly prompt in favour of tasks that require contextualised application. A graduate who has learned to apply concepts to a specific client, a specific organisation, a specific set of constraints, is a graduate who is ready for professional practice. A graduate who has learned to produce well-structured responses to abstract questions is a graduate who can pass assessments. The former is what employers are paying for. The distinction matters.

Timely feedback loops.

Feedback that arrives fast enough to act on, tied to the next stage of the work. This develops perhaps the most underrated professional capability of all: the ability to receive critical input, process it constructively, and improve in response to it. The graduates who struggle most in their first years of professional life are rarely those who lack technical knowledge. They are those who have never had to work under the discipline of iterative feedback —who treat evaluation as a verdict rather than an input. Assessment design can change that.

The bottom line

You don’t fix AI by designing it out. You fix it by designing tasks where understanding is the only viable path—and where the process of demonstrating that understanding builds the capabilities graduates will use every day.

That reframe matters because it changes what universities are actually producing. The graduate who has learned to think clearly, explain their reasoning under pressure, manage work iteratively, apply concepts in context, and incorporate feedback constructively—that graduate is prepared for the world they’re entering. Not because those skills were taught in a lecture, but because the assessment system demanded them.

Right now, many assessment systems are demanding the wrong things. They are demanding compliance with restrictions that will be obsolete in two years. They are demanding outputs that can be generated without understanding. They are demanding performance under artificial conditions that bear no resemblance to any professional environment.

The employers I speak with are not asking for graduates who avoided AI. They are asking for graduates who can think, communicate, and adapt. Those capabilities are not produced by restriction. They are produced by design. And the institutions that understand that distinction—and act on it—will produce graduates that the market actually wants.